Sprint v. Cogent, some clarity & facts

Having skimmed the Sprint / Cogent threads, I saw multiple errors and lots of really bad guesses. Instead of replying individually, I thought I would sum up a few facts so everyone was on the same page. This way when we run off into another 100 post thread, we can at least -start- from reality (although I would bet serious cash on long odds we will diverge from it soon enough).

1. Neither Sprint nor Cogent have transit
Both Sprint & Cogent are transit-free networks. (Notice how I carefully avoided saying "tier one"?) Whether one or both _should_ have transit is not a fact, and therefore outside the scope of this e-mail, but that neither have transit today is a fact. (And please don't tell me how Network X has 100 Mbps of transit in Sri Lanka because they are too lazy to lease undersea cable. If you don't understand what I am saying here, stop reading now.)

2. The Internet cannot "route around" de-peering
I know everyone believes "the Internet routes around failures". While occasionally true, it does not hold in this case. To "route around" the "failure" would require transit. See item #1.

3. Standard transit contracts do not guarantee full connectivity
If you are a Cogent customer, it is very unlikely your contract will allow you SLA or other credits for not being able to reach Sprint unless you negotiated something special. I doubt Sprint's standard contract is much different. Transit contract SLAs end at AS boundaries. This is because Network A has no control over Network B and therefore will not give credit if Network B fails. Of course, you can still sue, threaten to terminate, etc., but the letter of the contract almost certainly says nothing about packets going beyond your transit provider's ASN.

4. There is a reason behind ratios which has nothing to do with telco "sender-pays"
Hot potato routing + very poor ratios puts much more of the cost on the receiving network. This is a valid, logical, and costly concern for receiving networks. The concern can be alleviated by cold-potato routing through accepting MEDs, anycast, CDN, and other technologies; to which the receiving network may say they cannot send proper MEDs, etc. Whether the problem can or should be worked through is not a fact, though. That this issue has nothing to do with telco "sender-pays" mentality is. (Of course, the telcos might still have that mentality, but that doesn't change the facts.)

5. Cogent has been disconnected several times
Cogent has been de-peered (e.g. Teleglobe, L3, Sprint) and/or performed de-peering themselves (e.g. Telia) multiple times. Cogent has been disconnected from another network more times & for longer (in each instance?) than every other transit free network combined for the last decade. (In fact, if memory serves, for the history of the Internet - but I'm not quite sure enough to guarantee it as fact.) Cogent has also de-peered many non-transit-free networks, at least sometimes without even notifying the peer prior to disconnection. Whether that makes Cogent the bully-er or the bully-ee is not fact, so I will not comment on that here.

There are probably other errors I missed while skimming the longer posts. But this should get us started on a good, clean, factual footing for future flights of fancy.

How do you explain Cogent's arrangement with NTT (AS 2914)? If it's
not transit, what is it?

Does Akamai have peering arrangements with Cogent directly?

Paul

* Patrick W. Gilmore:

1. Neither Sprint nor Cogent have transit
Both Sprint & Cogent are transit-free networks. (Notice how I
carefully avoided saying "tier one"?) Whether one or both _should_
have transit is not a fact, and therefore outside the scope of this e-
mail, but that neither have transit today is a fact. (And please
don't tell me how Network X has 100 Mbps of transit in Sri Lanka
because they are too lazy to lease undersea cable. If you don't
understand what I am saying here, stop reading now.)

2. The Internet cannot "route around" de-peering
I know everyone believes "the Internet routes around failures". While
occasionally true, it does not hold in this case. To "route around"
the "failure" would require transit. See item #1.

Out of curiosity, what would happen if one of the parties got transit
from a business POV? Not just in this particular case, but in
general.

Doesn't this work because they are so large that any such arrangement
would immediately threaten traffic ratios at the (transit-free)
transit provider?

3. Standard transit contracts do not guarantee full connectivity

If this were true, why would end users (or, more generally, not
significantly multi-homed networks) buy transit from such networks?

Quite frankly, if any potential transit provider tried to make noises about
being able to *guarantee* full connectivity, I'd show him the door.

Consider the average length of an AS path. Now consider that your AS is
at one end, your transit provider is the next hop - and there's often 5 or 6
or more AS hops past that. And that potential transit provider has
absolutely *no* control over what some backhoe just did to connectivity
4 AS down the path...

For example, look at the traceroute from my desktop to where your mail
originated:

traceroute to 212.9.189.177 (212.9.189.177), 30 hops max, 60 byte packets
1 isb-6509-1.vl103.cns.vt.edu (128.173.12.1) 0.394 ms 0.712 ms 0.791 ms
2 isb-6509-2.po51.cns.vt.edu (128.173.0.5) 0.597 ms 0.681 ms 0.756 ms
3 isb-7606-2.ge1-1.cns.vt.edu (192.70.187.218) 0.740 ms 0.709 ms 0.687 ms
4 192.70.187.10 (192.70.187.10) 7.590 ms 7.610 ms 7.647 ms
5 te2-1--580.tr01-asbnva01.transitrail.net (137.164.131.177) 89.583 ms 89.618 ms 89.797 ms
6 llnw-peer.asbnva01.transitrail.net (137.164.130.30) 11.956 ms 9.450 ms 9.473 ms
7 ve5.fr3.iad.llnw.net (69.28.171.213) 17.243 ms 9.689 ms 17.443 ms
8 * * *
9 FRA-3-eth0-403.de.lambdanet.net (81.209.156.9) 99.266 ms 99.180 ms 99.163 ms
10 FRA-1-eth000.de.lambdanet.net (217.71.96.69) 98.342 ms 98.436 ms 98.283 ms
11 STU-3-pos330.de.lambdanet.net (217.71.96.82) 111.748 ms 111.764 ms 107.438 ms
12 bond0.border2.LF.net (212.9.160.73) 104.380 ms 104.404 ms 104.262 ms
13 em1.core.LF.net (212.9.160.65) 104.622 ms 104.761 ms 104.504 ms
14 dsl-gw.ispeg.de (212.9.161.26) 106.013 ms 105.999 ms 105.973 ms
15 dsl.enyo.de (213.178.172.64) 135.094 ms 136.729 ms 136.007 ms

Are you saying that you'd accept a contract where ispeg.de or LF.net are making
claims they can guarantee connectivity to AS1312 no matter what transitrail
is doing? (I admit being surprised - I was *expecting* the traceroute to
go through Level3 or Sprint, actually. When did lambdanet land in DE? :wink:

(And the real kicker - if transitrail burps, is ispeg or LF able to find us
via our Level3 or Sprint connections? Maybe, maybe not...)

1. Neither Sprint nor Cogent have transit
Both Sprint & Cogent are transit-free networks. (Notice how I carefully
avoided saying "tier one"?)

How do you explain Cogent's arrangement with NTT (AS 2914)? If it's
not transit, what is it?

So you are asking to have what was Cogent's arrangement with NTT?
That's like asking what was AOL's relationship with Level(3) when they
had their peering spat with Sprint. Just because a relationship
existed in the past does not mean the same relationship exists today.
  Accept the fact that both providers are transit free and move
forward.

Does Akamai have peering arrangements with Cogent directly?

Akamai are self declared peering sluts. So, yes, they have direct
peering arrangements with Cogent.

charles

Does Akamai have peering arrangements with Cogent directly?

Akamai are self declared peering sluts. So, yes, they have direct
peering arrangements with Cogent.

Hrm, so after I posted this, I looked a bit deeper into it and found:

3 vl3493.mpd03.jfk02.atlas.cogentco.com (154.54.5.226) 0.665 ms
0.670 ms 0.667 ms
4 te9-4.mpd01.jfk05.atlas.cogentco.com (154.54.26.62) 0.697 ms
0.686 ms 0.653 ms
5 gblx.jfk05.atlas.cogentco.com (154.54.11.138) 0.721 ms 0.717 ms 0.802 ms
6 te3-2-10g.ar4.nyc1.gblx.net (67.16.131.105) 0.986 ms 0.934 ms 1.181 ms
7 bandwidth-consulting.tengigabitethernet8-3.ar4.nyc1.gblx.net
(64.210.29.14) 0.696 ms 2.665 ms 0.700 ms
8 te-8-3.bbr1.ash1.bandcon.com (216.151.179.225) 7.187 ms 7.177 ms 7.073 ms
9 209.234.254.198 (209.234.254.198) 7.322 ms 7.141 ms 7.217 ms
10 84.53.144.71 (84.53.144.71) 7.053 ms 7.071 ms 7.016 ms

So, for at least for traceroute www.akamai.com, Coget is using their
peer (Global Crossing) to reach Akamai transit provider. Guess I
got caught up in the past myself, thinking "Cogent would never depeer
a slut like Akamai".

charles

1. Neither Sprint nor Cogent have transit
Both Sprint & Cogent are transit-free networks. (Notice how I carefully
avoided saying "tier one"?)

How do you explain Cogent's arrangement with NTT (AS 2914)? If it's
not transit, what is it?

I do not know, and neither do you. But I do know it is not "transit", at least not to Sprint.

It is trivial to prove to yourself if Cogent has transit. Find me any AS path in the global table showing "_TF1_TF2_174_", there "TF1" and "TF2" are the ASNs of two of the other 13 transit free networks. (Modulo a few leaked prefixes, which always seem to crop up. For instance, if a network has 40K prefixes in its cone, showing O(10) paths is not proof.)

This is a positive test - if you see it, you know they have transit, if you do not see it, you do not know they do not have transit. But combined with bifurcation when Sprint drops peering to Cogent, one can _know_ Cogent does not have full transit, or partial transit to Sprint. It is possible (although I personally believe unlikely) Cogent has partial transit to some other transit free network that you cannot see right now because their peering to that network is up and overriding the AS paths in the global table. But that doesn't matter to this discussion.

Does Akamai have peering arrangements with Cogent directly?

That is none of your business, not to mention completely irrelevant to the topic at hand as Akamai is neither a network nor transit free.

* Patrick W. Gilmore:

1. Neither Sprint nor Cogent have transit
Both Sprint & Cogent are transit-free networks. (Notice how I
carefully avoided saying "tier one"?) Whether one or both _should_
have transit is not a fact, and therefore outside the scope of this e-
mail, but that neither have transit today is a fact. (And please
don't tell me how Network X has 100 Mbps of transit in Sri Lanka
because they are too lazy to lease undersea cable. If you don't
understand what I am saying here, stop reading now.)

2. The Internet cannot "route around" de-peering
I know everyone believes "the Internet routes around failures". While
occasionally true, it does not hold in this case. To "route around"
the "failure" would require transit. See item #1.

Out of curiosity, what would happen if one of the parties got transit
from a business POV? Not just in this particular case, but in
general.

From a business perspective, one of the two parties would then be paying a third party to reach the other. In fact, a year ago this is exactly what was happening - Cogent bought partial transit from Verio to reach Sprint and AOL.

Neither believes this is in their best interest. I cannot tell you if that is true.

Doesn't this work because they are so large that any such arrangement
would immediately threaten traffic ratios at the (transit-free)
transit provider?

Obviously not since it was happening in the past. But you make a good point. Traffic from either of these networks is probably large enough to push at least one of the other transit-free networks over their peering ratios with someone else.

But probably not all. It is probable some transit free networks gets more traffic from Sprint than they send, so selling transit to Cogent would not hurt them. Not so sure any transit-free network pushes more to Cogent than they receive, but I cannot prove it.

And even if selling to Cogent would put them over their ratio requirements, perhaps they could negotiate a better settlement deal, so they get more from Cogent than they pay to Sprint.

Despite the fact I believe Cogent is heavy outbound to all other transit free networks, there are solutions that would allow a network to sell Sprint transit to Cogent.

Etc., etc.

It is a business problem, it has multiple business solutions.

3. Standard transit contracts do not guarantee full connectivity

If this were true, why would end users (or, more generally, not
significantly multi-homed networks) buy transit from such networks?

"If this were true"? "Why would end users [...] buy transit from such networks"?

Please show me a transit contract - just one - that guarantees connectivity beyond the transit AS boundary.

Put another way, since _every_ network does this, if you do not want to buy from 'such networks', you cannot buy transit.

Put another way, since _every_ network does this, if you do
not want to buy from 'such networks', you cannot buy transit.

Let's put it another 'nother way.
Would an end user get better connectivity by buying from a
reseller of transit? In other words, buying transit from
a network which also buys transit. Presumably up near the
top of the chain (Tier 1 vicinity), that transit reseller
has a lot of peering in place with other folks in the same
neighborhood (Tier 1 vicinity). But as long as a network
is a transit reseller (i.e. they buy transit), then they
are less likely to suffer from partition events caused
by fractious peering negotiations.

--Michael Dillon

frac*tious (frakshus)
adj.
1. Inclined to make trouble; unruly.
2. Having a peevish nature; cranky.

Also likely to cause your network having connectivity
to only a fraction of the Internet.

Quite frankly, if any potential transit provider tried to make
noises about
being able to *guarantee* full connectivity, I'd show him the door.

Let's not make the perfect the enemy of the good. All that's required is
that they promise to make a good faith effort to interconnect with anyone
else who makes a similar good faith effort.

Now we can all fight about whether Cogent and/or Sprint are making such a
good faith effort. :wink:

DS

Can anyone explain to me why end users find it so important to label carriers as "Tier 1" or "Tier 2"? The prevailing theory in the heads of prospective customers is that a "Tier 1" is somehow inherently better than a "Tier 2" (or lower), even though they don't quite understand the concepts behind why the "Tier" designation even exist(s/ed). These labels, at least to me, are no longer very relevant in today's internet world. In fact, would anyone agree that being a "Tier 1", as Cogent believes themselves to be, leaves that network in a very painful position when things like their frequent peering disputes happen?

For an NSP, it's obviously a "good thing" to be SFI-only, as in theory, it _should_ lower your costs. YMMV, as mentioned in a previous thread. However, what does it really matter to an end-user, especially if they are biased towards using "Tier 1" networks only? Why does a network who purchases transit give the impression to end users that that network's internet genitalia is somehow smaller than, say, Verizon or AT&T? I can see merit in touting the size and coverage of the actual network, but it's always been my understanding that this is not the true definition of the tiered system.

-evt

Patrick W. Gilmore wrote:

4. There is a reason behind ratios which has nothing to do with telco
"sender-pays"

There is an alleged reason.

Hot potato routing + very poor ratios puts much more of the cost on
the receiving network. This is a valid, logical, and costly concern
for receiving networks.

So what? So the argument is:

1) Your customers want to receive from my customers.

2) Receiving is more expensive.

3) Therefore you should pay me?

I want to send, and sending is cheap. Your customers want to do the
expensive receiving, not mine. My customers want to do the cheap sending.

The ratio argument is nonsense. If your customers want to receive mostly,
and receiving is expensive, they should pay you more to cover your higher
costs in receiving traffic. If my customers mostly want to send, and sending
is cheap, then I should pay less, since I want to do the cheap thing and you
want to do the expensive thing.

If we hold a convention in Idaho, the people who live near Idaho don't pay
money to the people who came from out-of-town. And if you live in Alaska,
get used to the fact that you're going to be paying more than your share of
transportation expenses to go to conventions. If you don't like it, don't
live in Alaska.

That this issue has nothing to do with telco "sender-
pays" mentality is. (Of course, the telcos might still have that
mentality, but that doesn't change the facts.)

You are probably right that it's not an old-fahsioned "sender-pays"
mentality. But it is complete nonsense. Your customers pay you to receive
their traffic, even if that's more expensive than other things they might
want to do.

Your customers pay you to carry their traffic across your network between
them and the next network in the line. There is no reason anyone else should
compensate you for doing this.

DS

From: michael.dillon@bt.com [mailto:michael.dillon@bt.com]
Sent: Monday, November 03, 2008 8:55 AM

Let's put it another 'nother way.
Would an end user get better connectivity by buying from a
reseller of transit? In other words, buying transit from
a network which also buys transit. Presumably up near the
top of the chain (Tier 1 vicinity), that transit reseller
has a lot of peering in place with other folks in the same
neighborhood (Tier 1 vicinity). But as long as a network
is a transit reseller (i.e. they buy transit), then they
are less likely to suffer from partition events caused
by fractious peering negotiations.

--Michael Dillon

Can anyone explain to me why end users find it so important to label carriers as "Tier 1" or "Tier 2"?

In my experience, end users generally don't know and almost never care. It's the sales people who talk about tiers.

Regards
Marshall

In a message written on Mon, Nov 03, 2008 at 01:26:14AM -0500, Patrick W. Gilmore wrote:

Having skimmed the Sprint / Cogent threads, I saw multiple errors and
lots of really bad guesses. Instead of replying individually, I
thought I would sum up a few facts so everyone was on the same page.
This way when we run off into another 100 post thread, we can at least
-start- from reality (although I would bet serious cash on long odds
we will diverge from it soon enough).

I notice a significant bit of information which has not been posted
in any of the posts I have read.

If I were to rank networks by the difficulty of getting settlement
free peering I believe Sprint would be at the top of the list. The
traffic volume they demand is pretty much higher than anyone else,
the ratio they demand is tighter than anyone else, and they demand
you meet them at Sprint POP's. Not friendly carrier neutral colos,
they are stuck in the Telco world of we run half the circuits to
you and run half the circuits to us. Never mind that only a few
carriers are even in a Sprint POP to deliver a circuit.

I don't necessarily fault Sprint for setting their requirements as
such; if I were them I may well do the same. Basically they have
decided they have enough peers and attempt to keep their requirements
high enough such that they need not ever add more.

However, looking at the issue from an external view, since I work for
neither Cogent nor Sprint, I do believe there are two fundamental
"fairness" questions:

1) Do all of Sprint's current peers meet Sprint's peering policy?

2) Due to the size and customer base of Sprint's network is it possible
   for anyone to meet their peering policy?

While these are rarely tested by anyone who could enforce them (save
the odd merger of mega players) they really are the interesting
questions. Several of the "good old boys" in the settlement free
club have fallen from where they once were, yet they don't get
depeered. If you have peers that no longer meet your current peering
policy and don't get depeered, how is that fair? Of course, the
way the industry is structured those arrangements are completely
hidden, so we'll never know if everyone meets the criteria or no
one does.

The second is also very interesting. Most networks came up with
their criteria (say, 2:1 ratio) 10, 15, or even 20 years ago. The
network was a very different place at that time, when a T1 was a
large circuit and interconnects were DS-3's. In particular, most
access was symmetrical; T1's, DS-3's, maybe ethernet (the 10M kind)
in a colo; and what's more, everyone was selling the same products.
There was no DSL, no cable modems, no EV-DO cards.

There have been two interesting developments in the industry since
that time, and the traditional criteria don't serve them well.
Asymmetrical access and specialization. Take a pure end user play
network, like most of the cable modem networks, and peer them with
any content player and the ratio is never going to be 1.5:1 or 2:1,
but they both get benefit from the arrangement. Many of the new
networks have realized this and adjusted their criteria, many of
the old players, like Sprint, may have morphed into one of these
new buckets, but not realized it yet.

This is an area where I believe our industry needs to grow up. We
like to operate on the assumption that if we have to jump over a
particular bar to peer with someone that policy has been applied
fairly across all players and is reasonable. However, there is no
way for anyone to verify that; short of the DoJ stepping in during
a merger. I am fairly sure that the requirements are not enforced
evenly by some players (no comment on Sprint or Cogent in this
case); that they use the fact that actual traffic volumes and ratios
are hidden to play poker with other networks.

While I'm generally pro business and think this is a good thing; I
believe it has reached a point in our industry where it is damaging
the Internet. The increase in peering spats to me is an indication
that the players are not looking out for stability, performance,
and insuring their customers are getting the access that they pay
for, but rather that they are interested in playing poker with each
other. I don't even think it's about the money, but rather about
the power.

The interesting thing is that there is a disruptive force on the
horizon. The IPv6 transition will change the peering landscape.
Traffic volumes and ratios will be moving to new providers as people
transition technologies. This is going to lead to quite a shake
up on the peering front I'm afraid, and I fear the result is going
to be great instability. The lack of transparency means that those
in power will try and bluff their way forward; the question is how
many people who have some visibility will call that bluff.

In all, this situation just reminds me of what I already knew. When
there is a depeering both networks are at fault. The fact that
these two players couldn't come to some commercially acceptable
terms to both of them speaks volumes about both.

David Schwartz wrote:

Your customers pay you to carry their traffic across your network between them and the next network in the line. There is no reason anyone else should compensate you for doing this.
  
What it all comes down to is that the majority of eyeballs are on "residential" connections that are relatively expensive to provide but for which are sold at a relatively low price (often 1/10th as much per megabit of capacity). Those eyeball ISPs cannot or will not charge their customers the full cost of "receiving" traffic so they want money from the more profitable content ISPs "sending" the traffic to offset their losses.

This is also one of the reasons eyeball ISPs want to stamp out P2P: both ends of the connections are on unprofitable lines and there is _nobody_ paying for the traffic. Just follow the money.

S

David Schwartz wrote:

The ratio argument is nonsense. If your customers want to receive mostly,
and receiving is expensive, they should pay you more to cover your higher
costs in receiving traffic. If my customers mostly want to send, and sending
is cheap, then I should pay less, since I want to do the cheap thing and you
want to do the expensive thing.

If it costs one party to an SFI agreement more than the other (total cost,
including intangibles) this makes the agreement less attractive, perhaps to the
point of inequitability. Where one party profits more from the agreement than
another, there is less incentive for the interconnection to be settlement-free.

There is no father figure standing there saying 'Party A and Party B must SFI
regardless of cost' - that decision is up to the relevant commercial minds
within Party A and Party B to carry out the required analysis and negotiate as
required.

Will

Patrick W. Gilmore wrote:

4. There is a reason behind ratios which has nothing to do with telco
"sender-pays"

There is an alleged reason.

Peering rations were first 'big news' when BBN wanted to de-peer Above.Net, Global Center, and Exodus in 1998. I spent a long time chatting with BBN's CTO about why BBN wanted to do this. I am convinced the facts are correct.

Perhaps more importantly, anyone who understands how BGP, fiber, routers, etc. work can figure this out for themselves without even talking to another network. Put another way, this is not a fantasy, supposition, bluster, etc.

What do you have to convince people otherwise?

Hot potato routing + very poor ratios puts much more of the cost on
the receiving network. This is a valid, logical, and costly concern
for receiving networks.

So what? So the argument is:

1) Your customers want to receive from my customers.

2) Receiving is more expensive.

3) Therefore you should pay me?

I don't remember saying that at all. Perhaps you should re-read my post.

I want to send, and sending is cheap. Your customers want to do the
expensive receiving, not mine. My customers want to do the cheap sending.

The ratio argument is nonsense. If your customers want to receive mostly,
and receiving is expensive, they should pay you more to cover your higher
costs in receiving traffic. If my customers mostly want to send, and sending
is cheap, then I should pay less, since I want to do the cheap thing and you
want to do the expensive thing.

The ratio argument is not nonsense. And fortunately, what you spout on NANOG has no effect on reality.

Your customers pay you to carry their traffic across your network between
them and the next network in the line. There is no reason anyone else should
compensate you for doing this.

<eyeball-network advocate>
Your customers pay you to deliver their traffic to my eyeballs. There is no reason I should compensate you for doing so.
</advocate>

The FACT is that a point-source sending traffic to distributed receivers combined with hot-potato routing puts more of the cost on the receiver. That fact is not in dispute, apparently even you agree.

From that fact, you can argue whether that is grounds for de-peering, settlements, etc. But the fact stands.

Also, please note no one is forcing you to pay anyone. Cogent decided not to pay. There is no law forcing them, Sprint is not holding a gun to Dave's head. But just like no one is forcing the sender to pay, no one is forcing the receiver to pay either.

Personally I think business problems have a business solution. For the ratio problems in 1998, Above.Net (and others, probably), agreed to carry the traffic and deliver it closer to BBN's eyeballs, thereby shifting the majority of the cost to AN. Dave (Rand this time, not Schaeffer) actually preferred it that way, saying he trusted his network more than BBN's and AN's customers pay him for quality. Hrmm, sounds like just the opposite of how you treat your customer's traffic....

* Stephen Sprunk

What it all comes down to is that the majority of eyeballs are on
"residential" connections that are relatively expensive to provide
but for which are sold at a relatively low price (often 1/10th as
much per megabit of capacity). Those eyeball ISPs cannot or will not
charge their customers the full cost of "receiving" traffic so they
want money from the more profitable content ISPs "sending" the
traffic to offset their losses.

Another point worth mentioning is that the traffic is going to flow
between those two ISPs _anyway_. Therefore, in many cases the only
ones to profit from them not reaching a peering agreement
(settlement-free or not) is their upstream(s), who is probably
delighted to be able to charge them both for the transit traffic.

Regards,

Another point worth mentioning is that the traffic is going to flow
between those two ISPs _anyway_.

I believe the events of 2-3 days ago disproves your assertion.

Therefore, in many cases the only
ones to profit from them not reaching a peering agreement
(settlement-free or not) is their upstream(s), who is probably
delighted to be able to charge them both for the transit traffic.

Again, supposed facts not in evidence.

I mentioned in the thread earlier that it is entirely possible Eyeball Network saves money by turning down peering and paying a transit provider to deliver the packets where Eyeball Network wants. Fiber, routers, IX ports, engineers, etc. are all expensive. Transit these days is not.

Doesn't mean Eyeball Network actually does save money. Just means you don't know either way.

In a message written on Mon, Nov 03, 2008 at 10:40:46AM -0500, Patrick W. Gilmore wrote:

The FACT is that a point-source sending traffic to distributed
receivers combined with hot-potato routing puts more of the cost on
the receiver. That fact is not in dispute, apparently even you agree.

s/more of the cost/more of the network transport cost/.

Having been at AboveNet when several providers tried to tell us it
was unfair that we had data centers near exchanges and it cost them
a lot of money to put in access to the end user I simply pointed
them to our stock reports where we were spending hundreds of millions
of dollars building data centers, and our customers were further
buying hundreds of millions of dollars of servers to serve up the
content.

You are correct that hot potato routing makes more of the bits flow
on the "eyeball" network. This of course can be mitigated by using
alternate routing strategies; for instance AboveNet actively
encouraged other providers to send us Meds and in some cases
deaggregate to us such that we could carry the bits on our network.
However I am skeptical when looking at the total system cost that
the costs are as disproportionate as you suggest.

It's great to build an eyeball network, but unless someone invests
in data centers and servers it's not going to get to any content.
The view that network transport cost is the only interesting figure
is a historical artifact of the days of $5000/megabit transit.
There was a time it was more expensive to carry bits across the
country than to build a data center; if anything I think it's now
the exact opposite. The data center providers are actually taking
more of the costs than many end user ISP's are due to the falling
price of transport and the relatively static price of real estate,
generators, air conditioners, and the like.